1. Technical Field
The present invention relates to a cluster system and a communication method in the cluster system and, more particularly, to a cluster system based on a parallel computing framework and a communication method in the cluster system.
2. Description of the Related Art
Recently, Open Computing Language (OpenCL) has attracted attention as a representative parallel programming model.
OpenCL is an open general-purpose parallel computing framework, and enables programs operating on a multi-platform (e.g., a plurality of central processing units (CPUs) or graphics processing units (GPUs)) to be created.
OpenCL can operate on various types of hardware that are manufactured by various hardware manufacturers. Many manufacturers are fabricating frameworks suitable for their hardware based on OpenCL.
In connection with this, in a conventional cluster system including a plurality of nodes, OpenCL operates at a host node, and other nodes frequently communicate the host node in order to perform computation. This deteriorates the performance of the overall cluster system.
That is, one of the nodes that form a common cluster system may become a host node, and the host node may execute OpenCL. This host node may copy data into computation devices, and may receive the results of computation that has been performed by each of the computation devices. In this processing process, the host node should continue to communicate with other nodes present within the cluster system. In this case, a problem arises in that the performance of the cluster system is deteriorated because a bottleneck phenomenon attributable to the communication occurs as the number of nodes within the cluster system increases.
Korean Patent Application Publication No. 10-1998-0036294, which is a related art document, proposes a data distribution and communication method in a shared memory system, in which the distribution of data in a shared memory form and the distribution of data in a distributed memory form are employed in a mixed manner by utilizing proposed HPCC functions using MPI functions, so that the distribution of data in a shared memory form is performed within one node and the distribution of data in a distributed memory form is performed outside the node. However, this technology does not solve the aforementioned problem.
Accordingly, there is a need for a technology that is capable of solving the aforementioned problem.
Meanwhile, the above-described related art corresponds to technical information that has been possessed by the present inventor in order to contrive the present invention or that has been acquired in the process of contriving the present invention, and cannot be necessarily viewed as technology well known to the public before the filing of the present invention.